[hibernate-dev] Cache region names are not prefixed in 5.3

[Sanne Grinovero]

On 17 May 2018 at 12:09, Radim Vansa <rvansa at redhat.com> wrote:
> I basically agree with Sanne here that having the prefix isolated opens
> space for performance improvements, though if certain call is prefixed,
> RegionFactory can always drop that prefix. The important thing is to mention
> if the region name is prefixed or not in javadocs. I would also prefer if
> *all* region names that RegionFactory is supposed to access followed the
> same strategy (prefixed/not-prefixed).
>
> 5.1 included method
>
>     QueryResultsRegion buildQueryResultsRegion(String regionName, Properties
> properties)
>
> where StandardQueryCache did the prefix for us, the change in 5.3 to
>
>     QueryResultsRegion buildQueryResultsRegion(String regionName,
> SessionFactoryImplementor sessionFactory)
>
> does not indicate that the prefix won't be there and the javadoc is missing
> completely.
>
> Also, DomainDataRegionConfig.getRegionName() has no javadoc and
> RegionFactory does not add the prefix.
>
> @Steve could you please amend the javadoc and confirm if RF should prefix
> region names?
>
> @Sanne while cache manager per deployment is an obvious way to distinguish
> regions at deployments, CM holds some heavy resources (e.g. threads) - so I
> while we are supposed to scale number of caches up to thousands (and we've
> fixed some problems that arise when you have for example ~3k caches), ATM
> you're not supposed to scale CMs. And I don't think that we'll focus in this
> direction since the bright future lies in microservices :)

Right, good points. It's a possible optimization I'd like to see
eventually but it's not trivial to get there yet.

Yet assuming a microservices scenario, this does become trivial to
benefit from: the container knows there's a single deployment, a
single CM. So no need to isolate them at all, just need to possibly
isolate multiple PUs defined in the same service.

So it will be easy to run hundreds of isolated microservices on the
same physical CPU and kill each other via process contention :P

>
> Radim
>
>
> On 05/17/2018 12:23 PM, Sanne Grinovero wrote:
>>
>> On 17 May 2018 at 11:11, Sanne Grinovero <sanne at hibernate.org> wrote:
>>>
>>> I think this is the RegionFactory's responsibility, as Hibernate ORM
>>> alone can't know if this is necessary.
>>>
>>> Prefixing is one of many options to isolate caches; a Cache technology
>>> might wish to use a different approach by implementing a custom
>>> `org.hibernate.cache.spi.CacheKeysFactory`.
>>
>> Hum, I regret how I wrote the above paragraph.
>>
>> I actually meant that a Cache technology could implement isolation in
>> many different ways.
>> Using a custom `CacheKeysFactory` is just an example, there are many
>> other options as well. In fact, I believe a good choice for
>> application servers would be to have an independent CacheManager for
>> each deployment. Then it can safely inspect the deployment, see if
>> there are multiple Persistence Units using the same caching
>> technology, and implement further isolation only if necessary.
>>
>> These thoughts are a consequence of a chat I had with Paul Ferraro
>> from the WildFly team, as he mentioned the size of the keys becoming
>> problematic with all the additional prefixing they need to apply. I
>> hope this will make it possible to e.g. create an "as small as
>> possible" unique identifier for the deployments to replace the
>> deployment name during serialization (to identify the CacheManager)
>> followed by a numeric id indexing the PUs within the deployment - or
>> nothing altogether in case of single PUs.
>>
>> Of course, I don't expect that to be needed right away; the
>> RegionFactory could just do good old prefixing for now but I hope
>> we'll explore such improvements in the near future.
>>
>> Thanks,
>> Sanne
>>
>>> Not least the server / deployer might be able to hint the Cache
>>> provider to tell if isolation is at all necessary.
>>>
>>> In conclusion, by having Hibernate ORM not messing with prefixes
>>> allows other technologies to implement more efficient solutions. Our
>>> own code also ends up being more efficient by not needing to add a
>>> prefix during each and every access to the cache.
>>>
>>> Thanks,
>>> Sanne
>>>
>>> On 17 May 2018 at 06:34, Gail Badner <gbadner at redhat.com> wrote:
>>>>
>>>> I see that cache region names are not being prefixed in 5.3.
>>>>
>>>> EnabledCaching sets the TimestampsRegion region name
>>>> to TimestampsRegion.class.getName(), and sets the QueryResultsRegion
>>>> region
>>>> name to QueryResultsRegion.class.getName(). [1]
>>>>
>>>> Without a prefix, WildFly is failing intermittently when there are 2
>>>> persistence units with the query cache enabled due to:
>>>>
>>>> org.infinispan.commons.CacheConfigurationException: ISPN000453: Attempt
>>>> to
>>>> define configuration for cache org.hibernate.cache.spi.TimestampsRegion
>>>> which already exists
>>>>
>>>> Entity region names are not being prefixed either.
>>>>
>>>> Should they be prefixed by Hibernate or by the RegionFactory?
>>>>
>>>> Regards,
>>>> Gail
>>>>
>>>> [1]
>>>>
>>>> https://github.com/hibernate/hibernate-orm/blob/master/hibernate-core/src/main/java/org/hibernate/cache/internal/EnabledCaching.java#L80-L92
>>>> _______________________________________________
>>>> hibernate-dev mailing list
>>>> hibernate-dev at lists.jboss.org
>>>> https://lists.jboss.org/mailman/listinfo/hibernate-dev
>
>
>
> --
> Radim Vansa <rvansa at redhat.com>
> JBoss Performance Team
>